Interview with Gregor Blichmann
In the years since our founding, we have seen the ways in which AI implementation doesn’t work well: Development should not be oriented toward the projects of individual customers with specific problems. One-off solutions to such problems cannot be transferred to other projects, meaning that the time and money was not exactly invested sustainably. Instead, we now provide companies that want to get specific results from their data with the appropriate AI building blocks, enabling them to process their customer data without having to start from scratch each time. There are, of course, limits to this principle, but it has proven to be very practical and sustainable – not only economically, but also environmentally. If a new model must be trained for each AI project, the server requirements and energy needed for computation are enormous. The less training we require, the fewer resources are consumed to build the models and thus deliver solutions to the customer.
Imagine a project where documents need to be extracted. If we were to develop an AI solution specifically for this project, we would have to, for example, build a model that can recognize the document type, extract certain areas and read the corresponding text. In our approach, the tasks in this process are treated as modular building blocks. There is a defined model for each step of the work – for handwriting recognition, for example. This module is not, or only slightly, dependent on the specific project profile of each customer. We can use these types of trained models in all use cases where they are needed. We have trained models for specific small tasks according to this building-block principle, and we then arrange the individual modules using a workflow, which is based on the project requirements.
We have recognized that we can’t offer AI software to indi- vidual customers with the expectation that we will sell it and be done with it. The implementation of AI follows certain workflows that are established within the company. These workflows evolve over time – also in the sense that the model quality improves with the trained elements. Trust must be es- tablished with clients. We work with transparent benchmarks and tell them, for example: We generated a representative data set and analyzed it with the models; 80 percent of the documents were recognized, and the data was read with 95-percent accuracy. Customers can then give us feedback as to whether these values are satisfactory for them. In the end, if the results are right, it doesn’t matter if the AI solution is a black box. But we must first get to the stage where they trust this black box.
What types of use cases are companies approaching you with?
Our core business is the development of AI products for the automation of business processes, like automated document processing. For example, automated capture and processing of forms for ordering or inventory, medical history forms in hospitals, invoices, delivery bills, service tickets, construction plans or pure text in the broadest sense.
What is the impact of automation on companies?
One of our customers is a manu- facturer of orthopedic products, to take one example. Individual orders make up a significant portion of its business. Patients have their legs or arms measured individually in a health care supply store and these measurements are then transmit- ted to the supplier, which must manufacture and ship the required product within a short period of time. Such a process has an extremely tight timeline and a high degree of individualization. Everything has to fit perfectly as well. Up to 40 measurements along with 40 to 80 configuration options must be considered for each order. To date, up to 80 percent of the orders are processed through a form sent in by email or fax. Our customer receives several thousand order forms each day, often filled out by hand. It used to be the case that around 40 employees would type in the forms in two shifts each day. But it was growing increasingly difficult for our client to find employees for this task, and that contravened the company’s growth plans. They wanted to process more orders each day but doing so wasn’t possible with the processes in place. Our task was to recognize the orders in a partially or fully automated way and to enter them into the system.
When automation processes are introduced, there is always concern that jobs will be lost as a result. How did you handle those fears?
We have to bear in mind that our society is facing a demographic change in which a great many people will be leaving the workforce in the next 10 years. At the same time, the number of documents, information and data that need to be processed will steadily increase. AI is one solution for addressing this everwidening gap between available workforce and the need for processing data. This isn’t about taking away people’s jobs, but about addressing this problem. In addition, AI should make work easier, and employees should be deployed where AI cannot provide assistance.
When you attempt to automate a process, it’s not enough to just use AI. A smooth process flow hinges on the specialized knowledge and experience of employees. For example, employees have known customer X for a long time. For years, when this customer orders something, they have always added an extra 10 percent to the specified dimensions. This is because they know from experience that this customer’s orders are always too tight, and without adjustment, he will change them later, anyway. For an automated process to be carried out successfully, such nuances must also be included in the target system. Mere data extraction isn’t enough. As such, along with the AI system, we also provide our customers with an editor they can use to define rules themselves. It allows you to specify in the system: If a specific customer number and the following attribute is recognized, then please delete that value or add 10 percent to all measurements, for example. We have built the tool in a way that our customers can make such edits themselves and consider what the system is still missing. In this way, they can gradually incorporate the existing specialist knowledge of individual employees into the system, bit by bit, to steadily increase the degree of automation.
A workshop was held with employees from the customer service department – those people who had previously typed in the data by hand – to explain how the rules could be adapted. They now maintain the system under the supervision of the head of customer service and an IT manager. But the rules come from the employees. The focus of their work is also shifting as a result of the use of AI. They now spend their time checking and processing borderline cases that can’t be handled by the system and have to be dealt with by the employees – who, in turn, now have more time to devote individually to these cases.
You work together with the sustainable data center Cloud&Heat. How important is it to consider resource conservation already in the development phase of AI?
There is a strong interaction between economic and environmental factors. For us, a key question is how to achieve the lowest possible computing time. Can I reduce the computing time simply by the choice of the architecture of my AI model and the software behind it? This is a very simple and important factor in being environmentally sustainable, because it means we use less electricity. And it has the economic advantage that it costs less. We also can’t forget about the hardware that is necessary for the calculations. The construction of the hardware already generates a large carbon footprint. We need to train our models on GPUs. If I need 100 GPUs, the footprint is correspondingly large. But if I choose the architecture of my model intelligently enough to compute a similar result on 10 GPUs, then I generate a much smaller carbon footprint. We are intrinsically motivated to optimize our models, but at the same time, there are a number of monetary incentives as well.